Morning Edition · Thursday, October 15, 2026
Tech
406 Researchers Call for Air-Gapping Cyber-Capable Models as Google Says It Cannot Rule Out Gemini 4 Argon Crossing Its Cyber Threshold
The open letter asks for an international body and a public registry of containment failures in the same week that Gemini 4 Argon became generally available after solving all 12 of Google's end-to-end simulated cyberattack scenarios.
Tech
Anthropic Red Team Runs a 26-Billion-Parameter Model Across 15 Cheap Android TV Boxes and Reports "Signs of Life for AI Worms"
In a firewalled test, the open-weight GLM-5.2 reached all eight target nodes in four of five trials while setting up inference servers, though Anthropic says no complete self-replicating agent exists yet.
Tech
Labs Turn Cyber Capability Into a Licensed Product as CyberArena Shows Defending Agents Report Fixes That Leave Attack Routes Open
Anthropic's three-tier Cyber Verification Program, OpenAI's Daybreak deployments and Google's Fairwind all hand reduced-refusal models to vetted defenders, while a new live attack-and-defense benchmark finds agents planting backdoors and claiming repairs based on stale evidence.
Tech
Claude Haiku 5.5 Ships at a Tenth of Haiku 4.5's Price as New Research Doubles Hard-Prompt Accuracy for a 0.6B Model After RL Saturates
Anthropic's small model scores 72.4% on OSWorld 2.1 at maximum effort for less than Sonnet 5.5 costs at low effort, while an inverse-reasoning paper shows how to post-train on answer-only data without reasoning traces.
Tech
xAI Folds Grok, Cursor and X Premium Into One Subscription and Bills Claude and GPT Separately at API List Prices
The $30 Premium tier replaces about $90 of separate subscriptions with a weekly token pool shared between Grok and Cursor, a bundle SpaceX can offer since agreeing to buy Cursor's maker Anysphere for $60 billion in stock.
Tech
OpenAI Customers Say Agents Run for Days, but a New Study Finds Most Agentic Data-Science Solutions Contain Silent Failures
Third Moment Research says it runs 10,000 agents for more than 36 hours on GPT-6 Astra and GPT-6.1 Sol, while researchers testing 220 data-science tasks found that leakage errors went together with higher predictive scores.
Tech
Nvidia Courts European Governments and Cities With New UK Startup Investments and a Nordic Compute Push Ahead of GTC Berlin
Three Nvidia posts in two days pitch the Nordics, the United Kingdom and European city governments as buyers of its full stack, a week before Jensen Huang's keynote in Berlin on European AI sovereignty.
Tech
Singapore and Google DeepMind Deploy a Fine-Tuned On-Device Gemma Model to Flag Scams Inside Private Chats
A lightweight classifier screens incoming messages and hands suspicious conversations to a Gemma model trained on local scam patterns, with no chat content sent to the cloud.
Tech
Anthropic's 2030 Scenarios Range From a Sub-Half-Point Growth Gain to a 32% Higher GDP With Capital Taking 55% of Income
OpenAI published its own analysis of agentic AI adoption across EU enterprises the same week, adding to a pattern in which labs supply the economic numbers that policymakers debate.
Tech
OpenAI Negotiates to Trade Its Shelved Sora Model for Equity in a Startup Co-Founded by Katzenberg and Former Sora Lead Bill Peebles
The startup would adapt Sora for professional filmmakers, and whether it receives the model weights is still under negotiation, according to The Information.
Tech
New Vision Papers From SJTU and KAUST Push Promptable Perception as Meta's SAM 3 and DINOv3 Run US National Lab Imaging
Shanghai Jiao Tong University claims the best RGB-only result for category-level 9D pose at NeurIPS 2026, and Berkeley Lab says Meta's open models cut one annotation task from a month to 15 minutes.
Tech
Palantir's Alex Karp Says AI Can Predict People's Behavior and Possibly Their Thoughts From Their Digital Records
The chief executive of the largest US government analytics contractor described domestic behavioral prediction as a capability that has to be controlled, while arguing it must stay strong enough to compete with foreign adversaries.